SoilSpecLLM is a soil VNIR-SWIR spectral learning framework for 400-2500 nm reflectance modeling. The framework aligns soil-property semantics with continuous spectral signals and supports two tasks:
- spectral prediction: predict missing/following spectral segments from preceding bands
- spectral generation: condition spectral synthesis on soil property descriptions
This repository version is organized for paper code release and currently focuses on the spectral prediction pipeline.
The method is designed for global-scale soil spectral data and targets:
- accurate full-curve reconstruction/prediction
- robust cross-sensor generalization
- physically consistent absorption behavior around 1.4, 1.9, and 2.2 um
- Soil spectrum
- Large language model
- Diffusion model
Soil Prediction/: complete spectral prediction code package used for this paper release.Soil Generation/: diffusion-based spectral generation code package for soil spectrum synthesis.